ChainThink reports: In response to efforts by Washington and Anthropic to cut off China’s access to frontier models by blocking APIs, Kevin S. Xu, former Head of Internationalization Strategy at GitHub and founder of Interconnected Capital, stated that adversarial distillation is merely a desperate shortcut adopted by some independent Chinese labs due to data scarcity, and blocking APIs cannot halt China’s overall progress in AI.
DeepSeek, Moonshot, and MiniMax, all independent labs without group ecosystem support, face a critical shortage of high-quality post-training data, such as reasoning steps.
Backed by major corporate labs such as Alibaba (Qwen), ByteDance (Seed), or Xiaomi, and possessing vast proprietary scenario data comparable to that of Google and Apple, these companies do not rely on distillation. Even restrictive policies can only cause short-term setbacks for independent labs and cannot undermine the foundation of China’s major tech corporations.
The much-talked-about Chinese "data advantage" is a misconception; in terms of high-quality annotated and evaluation data required for training cutting-edge large models, China not only lacks an advantage but also suffers from a severe shortage of mature commercialized data supply chains like Scale AI or Surge.
Due to the poor quality of domestic data service providers, independent labs have adopted API distillation as a cost-effective data acquisition strategy. The data labeling industry is a business model issue with low entry barriers, not a technical flaw, and the supply-demand gap in China can be easily filled.
In the long term, the theoretical upper limit of a distilled student model cannot surpass its teacher. However, Chinese developers will ultimately break through this limit and design large models that outperform their mentors. The U.S. blockade policy is not only ineffective but may also prematurely sever the theoretical constraints that could have locked Chinese models beneath the "student" ceiling.